bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.
1 Early Tetrapodomorph Biogeography: Controlling for Fossil Record 2 Bias in Macroevolutionary Analyses 3 4 Jacob D. Gardnera,1, Kevin Suryab,1, and Chris L. Organa* 5 6 a Department of Earth Sciences, Montana State University, Bozeman, MT 59717, USA 7 b Honors College, Montana State University, Bozeman, MT 59717, USA 8 9 * Corresponding author at: Department of Earth Sciences, Montana State University, Bozeman, 10 MT 59717, USA. E-mail address: [email protected] (C. L. Organ) 11 12 1 Contributed equally to this work 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 1
47 ABSTRACT
48 The fossil record provides direct empirical data for understanding macroevolutionary patterns
49 and processes. Inherent biases in the fossil record are well known to confound analyses of this data.
50 Sampling bias proxies have been used as covariates in regression models to test for such biases.
51 Proxies, such as formation count, are associated with paleobiodiversity, but are insufficient for
52 explaining species dispersal owing to a lack of geographic context. Here, we develop a sampling
53 bias proxy that incorporates geographic information and test it with a case study on early
54 tetrapodomorph biogeography. We use recently-developed Bayesian phylogeographic models and
55 a new supertree of early tetrapodomorphs to estimate dispersal rates and ancestral habitat locations.
56 We find strong evidence that geographic sampling bias explains supposed radiations in dispersal
57 rate (potential adaptive radiations). Our study highlights the necessity of accounting for geographic
58 sampling bias in macroevolutionary and phylogenetic analyses and provides an approach to test
59 for its effect.
60 Keywords: sampling bias, fossil record, biogeography, phylogenetics, macroevolution, tetrapod
61 water-land transition
62
63
64
65
66
67
68
69 bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 2
70 1. Introduction
71 Our understanding of macroevolutionary patterns and processes are fundamentally based
72 on fossils. The most direct evidence for taxonomic origination and extinction rates come from the
73 rock record, as do evidence for novelty and climate change unseen in data sets gleaned from extant
74 sources. There are no perfect data sets in science; there are inherent limitations and biases in the
75 rock record that must be addressed when we form and test paleobiological hypotheses. For instance,
76 observed stratigraphic ranges of fossils can mislead inferences about diversification and extinction
77 rates (Raup and Boyajian, 1988; Signor and Lipps, 1982). Observed species diversity is also known
78 to increase with time due to the preferential preservation and recovery of fossils in younger
79 geological strata—referred to as "the Pull of the Recent" (Jablonski et al., 2003). Large and long-
80 surviving clades with high rates of early diversification tend to result in an illusionary rate slow-
81 down as diversification rates revert back to a mean value—referred to as “the Push of the Past”
82 (Budd and Mann, 2018). Paleobiologists test and account for these biases when analyzing
83 diversification and extinction at local and global scales (Alroy et al., 2001; Benson et al., 2010;
84 Benson and Butler, 2011; Benson and Upchurch, 2013; Benton et al., 2013; Foote, 2003; Jablonski
85 et al., 2003; Koch, 1978; Lloyd, 2012; Sakamoto et al., 2016a, 2016b). These bias-detection and
86 correction techniques include fossil occurrence subsampling (Alroy et al., 2001; Jablonski et al.,
87 2003; Lloyd, 2012); correcting origination, extinction, and sampling rates using evolutionary
88 predictive models (Foote, 2003); the use of residuals from diversity-sampling models (Benson et
89 al., 2010; Benson and Upchurch, 2013; Sakamoto et al., 2016b); and the incorporation of sampling
90 bias proxies as covariates in regression models (Benson et al., 2010; Benson and Butler, 2011;
91 Benton et al., 2013; Sakamoto et al., 2016a). Benton et al. (2013), studying sampling bias proxies,
92 demonstrated that diversity through time closely tracks formation count (Benton et al., 2013). bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 3
93 However, case studies in England and Wales suggest that proxies for terrestrial sedimentary rock
94 volume (such as formation count) do not accurately explain paleobiodiversity, particularly if the
95 fossil record is patchy (Dunhill et al., 2014a, 2014b, 2013). Marine outcrop area and
96 paleoecological-associated facies changes are, however, associated with shifts in paleobiodiversity
97 (Dunhill et al., 2014b, 2013). Moreover, Benton et al. (2013) argue that the direction of causality
98 between paleobiodiversity and formation count is unclear; there may be a common cause to explain
99 their covariation, such as sea level (Benton et al., 2013). Nonetheless, formation count is a widely-
100 used sampling bias proxy in phylogenetic analyses of macroevolution (O’Donovan et al., 2018;
101 Sakamoto et al., 2016a, 2016b; Tennant et al., 2016a, 2016b). The advent of computational
102 modeling approaches, particularly phylogenetic comparative methods, has made it easier to
103 include proxies, like formation count, into models. Additional sampling bias proxies used in these
104 studies include occurrence count, valid taxon count, and specimen completeness and preservation
105 scores. Absent from these proxies is geographic context, which could confound many types of
106 macroevolutionary analyses.
107 Despite advancements made in understanding the origin and evolution of early
108 tetrapodomorphs, biogeographical studies are hindered by the incompleteness of the early
109 tetrapodomorph fossil record. For example, “Romer’s Gap” represents a lack of tetrapodomorph
110 fossils from the end-Devonian to mid-Mississippian, a period crucial for understanding early
111 tetrapodomorph diversification. Recent collection efforts recovered tetrapodomorph specimens
112 from “Romer’s Gap”, suggesting that a collection and preservation bias explains this gap (Clack
113 et al., 2017; Marshall et al., 2019). In addition, a trackway site in Poland demonstrates the existence
114 of digit-bearing tetrapodomorphs 10 million years before the earliest elpistostegalian body fossil,
115 showcasing the limitation of body fossils to reveal evolutionary history (Niedźwiedzki et al., 2010). bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 4
116 A recent study by Long et al. (2018) leveraged phylogenetic reconstruction of early
117 tetrapodomorphs to frame hypotheses about the origin of major clades, as well as their dispersal
118 patterns, including the hypothesis that stem-tetrapodomorphs dispersed from Eastern Gondwana
119 to Euramerica. However, this study did not use phylogenetic comparative methods to estimate
120 ancestral geographic locations or to model dispersal patterns.
121 Here, we present a phylogeographic analysis of early tetrapodomorphs. Our goals are: 1)
122 to construct a phylogenetic supertree of early tetrapodomorphs that synthesizes previous
123 phylogenetic reconstructions; 2) to estimate the paleogeographic locations of major early
124 tetrapodomorph clades using recently-developed phylogeographic models that account for the
125 curvature of the Earth; and 3) to test for the influence of geographic sampling bias on dispersal
126 rates. Our results indicate that geographic sampling bias substantially confounds analyses of
127 dispersal and paleogeography. We conclude with a discussion about the necessity of controlling
128 for fossil record biases in macroevolutionary analyses.
129 2. Materials and Methods
130 2.1. Nomenclature
131 Tetrapoda has been informally defined historically to include all terrestrial vertebrates with
132 limbs and digits (Laurin, 1998). Gauthier et al. (1989) first articulated a phylogenetic definition of
133 Tetrapoda as the clade including the last common ancestor of amniotes and lissamphibians. This
134 definition excludes stem-tetrapodomorphs, like Acanthostega and Ichthyostega. Stegocephalia
135 was coined by E.D. Cope in 1868 (Cope, 1868), but was more recently used to describe fossil taxa
136 more closely related to tetrapods than other sarcopterygians. A recent cladistic redefinition of
137 Stegocephalia includes all vertebrates more closely related to temnospondyls than Panderichthys
138 (Laurin, 1998). Here, we use the definitions of Laurin (1998) for a monophyletic Stegocephalia bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 5
139 and of Gauthier et al. (1989) for Tetrapoda, which refers specifically to the crown group. We use
140 Tetrapodomorpha to refer to all taxa closer to the tetrapod crown-group than the lungfish crown-
141 group (Ahlberg, 1998). We additionally use Elpistostegalia (= Panderichthyida) to refer to the
142 common ancestor of all stegocephalians and Panderichthys as well as Eotetrapodiformes to refer
143 to the common ancestor of all tristichopterids, elpistostegalians, and tetrapods (Coates and
144 Friedman, 2010).
145 2.2. Supertree
146 We inferred a supertree of 69 early tetrapodomorph taxa from five edited, published
147 morphological data matrices, focusing on tetrapodomorphs whose previously inferred
148 phylogenetic position bracket the water-land transition (Clack et al., 2017; Friedman et al., 2007;
149 Pardo et al., 2017; Swartz, 2012; Zhu et al., 2017). Since downstream analyses might be sensitive
150 to unequal sample sizes between taxa pre- and post-water-land transition, we did not include
151 several crownward stem-tetrapodomorphs from the original matrices (see Supplementary
152 Material). For each matrix, we generated a posterior distribution of phylogenetic trees using
153 MrBayes 3.2.6 (Ronquist et al., 2012b). In each case, we ran two Markov chain Monte Carlo
154 (MCMC) replicates for 20,000,000 generations with 25% burn-in, each with four chains and a
155 sampling frequency of 1,000. We used one partition, except for Clack et al.’s (2017) matrix, which
156 was explicitly divided into cranial and postcranial characters. To time-calibrate the trees, we
157 constrained the root ages and employed a tip-dating approach (Ronquist et al., 2012a). Tip dates
158 (last occurrence) were acquired from the Paleobiology Database (PBDB; https://paleobiodb.org/)
159 and the literature (see Supplementary Table 2). Root calibrations (minimum and soft maximum
160 age estimates) were collected from the PBDB and Benton et al. (2015). We also used the fossilized
161 birth-death model as the branch length prior (Didier et al., 2017, 2012; Didier and Laurin, 2018; bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 6
162 Gavryushkina et al., 2014; Heath et al., 2014; Stadler, 2010; Zhang et al., 2016). All pairs of
163 MCMC replicates converged as demonstrated by low average standard deviation of split
164 frequencies (<0.005; Lakner et al., 2008; see Supplementary Table 3).
165 Next, we used the five maximum clade credibility trees (source trees; Supplementary Fig.
166 1-10) to compute a distance supermatrix using SDM 2.1 (Criscuolo et al., 2006). We then inferred
167 an unweighted neighbor-joining tree (UNJ by Gascuel, 1997) from the distance supermatrix using
168 PhyD* 1.1 (Criscuolo and Gascuel, 2008). The UNJ* algorithm is preferable for matrices based
169 on morphological characters. Unlike most supertree methods, the SDM-PhyD* combination
170 produces a supertree with branch lengths. We rooted the supertree using phytools 0.6.60 (Revell,
171 2012) by adding an arbitrary branch length of 0.00001 to break the trichotomy at the basal-most
172 node in R 3.5.2 (R Core Team, 2018), designating the dipnomorph Glyptolepis as the outgroup.
173 We qualitatively compared the supertree topology with the published source trees and
174 Marjanović and Laurin's (2019) Paleozoic limbed vertebrate topologies. We also calculated
175 normalized Robinson-Foulds (nRF) distances (Robinson and Foulds, 1981) using phangorn 2.4.0
176 (Schliep, 2011) in R to assess the congruency of topologies. In each comparison, polytomies in the
177 supertree or the source tree were resolved in all possible ways using phytools. We then calculated
178 all nRF distances and took an average (see Supplementary Table 4). The supplementary materials
179 include a more detailed description of this approach.
180 2.3. Phylogeography
181 We obtained paleocoordinate data (paleolatitude and paleolongitude) for 63 early
182 tetrapodomorphs from the PBDB using the GPlates software setting (https://gws.gplates.org/). By
183 default, GPlates estimates paleocoordinates from the midpoint of each taxon’s age range. For 16
184 taxa that did not have direct paleocoordinate data in the PBDB, we searched for the geological bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 7
185 formations and geographic regions within the time range from which they are known and averaged
186 the paleolocations across each valid taxonomic occurrence in the PBDB. If the paleolocation of
187 the formation was not listed in the PBDB, we used published geographic locations of the
188 formations. This level of precision is adequate for world-wide phylogeographic analyses, such as
189 conducted here. Present-day coordinates for these geographic locations were obtained from
190 Google Earth and matched with PBDB entries that date within each taxon’s age range (see
191 Supplementary Table 5). Four additional taxa, Kenichthys, Koilops, Ossirarus, and Tungsenia, had
192 occurrences in the PBDB but the GPlates software could not estimate their paleocoordinates. For
193 Koilops and Ossirarus, we used all tetrapodomorph occurrences from the Ballagan Formation of
194 Scotland, UK—a formation in which these two taxa are found (Clack et al., 2017). For Kenichthys
195 and Tungsenia, we calculated paleocoordinate data from the GPlates website directly using the
196 present-day coordinates from the PBDB (https://gws.gplates.org/#recon-p). This approach did not
197 work for the 16 previously mentioned taxa (see Supplementary Table 5). We therefore obtained
198 paleocoordinate data from nearby entries in the PBDB. We excluded the following taxa from our
199 analyses due to the lack of data and comparable entries in the PBDB: Jarvikina, Koharalepis,
200 Spodichthys, and Tinirau. We excluded the outgroup taxon, Glyptolepis, in our analysis to focus
201 on the dispersal trends within early Tetrapodomorpha. We also excluded Eusthenodon and
202 Strepsodus because their high estimated dispersal rates—being reported from multiple
203 continents—masked other rate variation throughout the phylogeny and inhibited our downstream
204 analyses from converging on a stable likelihood. We do, however, discuss their geographic
205 implications in Section 4.
206 A model that incorporates phylogeny is crucial for paleobiogeographic reconstruction
207 because it accounts for both species relationships and the amount of evolutionary divergence bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 8
208 (branch lengths). Using continuous paleocoordinate data, rather than discretely-coded regions,
209 allows dispersal trends to be estimated at finer resolutions. Discretely-coded geographic regions
210 also limit ancestral states to the same regions inhabited by descendant species. However, standard
211 phylogenetic comparative methods for continuous data assume a flat Earth because they do not
212 account for spherically structured coordinates (i.e., the proximity of −179° and 179° longitudes).
213 Recently-developed phylogenetic comparative methods for modeling continuous paleocoordinate
214 data, implemented as the ‘geo’ model in the program BayesTraits V3, overcome this hurdle by
215 “evolving” continuous coordinate data on the surface of a globe (O’Donovan et al., 2018). The
216 model is implemented with a Bayesian reversible jump MCMC algorithm to estimate rates of
217 geographic dispersal and ancestral paleolocations simultaneously. To account for the spheroid
218 shape of the globe, the ‘geo’ model converts latitude and longitude data into three-dimensional
219 coordinates while prohibiting moves that penetrate the inside of the globe. Ancestral states, which
220 are converted back to standard latitude and longitude, are estimated for each node of the phylogeny.
221 The method includes a variable rates model to estimate variation in dispersal rate (Venditti et al.,
222 2011). The ‘geo’ model makes no assumptions about the location of geographic barriers or
223 coastlines, but a study on dinosaur biogeography found 99.2% of mean ancestral state
224 reconstructions to be located within the bounds of landmasses specific to the time at which they
225 occurred (O’Donovan et al., 2018). We ran three replicate independent analyses using the Bayesian
226 phylogenetic ‘geo’ model for 100 million iterations each with a 25% burn-in and sampling every
227 1,000 iterations. We estimated log marginal likelihoods using the Stepping Stone algorithm with
228 250 stones sampling every 1,000 iterations (Xie et al., 2011). We used Bayes factors (BF) to test
229 whether a variable rates model explained the data better than a uniform rates model. Bayes factors
230 greater than two are considered good evidence in support of the model with the greater log bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 9
231 marginal likelihood. We compared estimated rate scalars and ancestral states among the three
232 independent variable rates analyses to check for consistency in our results. Rates of dispersal were
233 estimated for each branch by dividing the average rate scalars by the original branch lengths
234 (scaled by time). We assessed the MCMC convergence of all analyses using Tracer 1.7 (Rambaut
235 et al., 2018).
236 To test for the effect of sampling bias on dispersal rates, we developed a sampling bias
237 proxy that incorporates geographic context: regional-level formation count. Formation counts are
238 meant to capture multiple biases: uneven global rock exposure, uneven fossil collection and
239 database efforts, and global variation in sediment deposition in environments conducive to
240 preservation. Stage-level (stage-specific) formation count represents the mean number of
241 formations, or distinct rock units, globally known to produce relevant fossils along each terminal
242 branch of a phylogeny. Following the protocol of Sakamoto et al. (2016) and O’Donovan et al.
243 (2018), stage-level formation counts are calculated by taking the average number of formations
244 known from each geological age across the globe that encompass the time period between the
245 taxon’s tip date and its preceding node. These average stage-level formation counts are weighted
246 by the proportion that each terminal branch length covers each geological age. For example, if a
247 terminal branch covers two geological ages (e.g., Frasnian and Famennian) at 30% and 70%,
248 respectively, then the stage-level formation counts from each geological age are weighted by those
249 proportions and then divided by the number of geological ages covered:
250
FrasnianCount × 0.3 + FamennianCount × 0.7 251 Stage − Level FormationCount = 2
252
253 Stage-level formation count is not informed by geography; it is a global metric. It is bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 10
254 therefore an inadequate proxy if bias has a strong geographic component (e.g., if the majority
255 of formations recorded are from a specific region or if few formations are exposed within a region).
256 The number of fossil-bearing geological formations, accounting for geographic distribution, is
257 expected to be an important confounding bias in the fossil record. We developed a proxy that
258 includes geographic sampling bias. Our approach breaks down stage-level formation count by
259 geographic region. To account for the arrangement of the continents during the Devonian,
260 Carboniferous, and Permian, we recognized five major regions: Northern Euramerica (including
261 Northeastern Eurasia and Central Asia), Southern Euramerica (North America, Greenland, and
262 Western Europe), Western Gondwana (South America and Africa), Eastern Gondwana (Antarctica,
263 Australia, and Southern Asia), and East Asia (e.g., China). For each branch in the phylogeny, we
264 used the average ancestral state and taxon paleolocation estimates to determine if the branch
265 crossed multiple geographic regions. The number of formations within this time window are
266 totaled for every region covered by the branch and then divided by the number of regions covered.
267 For example, if ancestral state estimates at node 1 and 2 are located in Eastern Gondwana and
268 Southern Euramerica, respectively, then the number of formations recorded in Eastern Gondwana,
269 Southern Euramerica, and the regions in between (i.e., Western Gondwana or Northern Euramerica
270 + East Asia) are counted for that geological age; this total is then divided by the number of
271 geographic regions covered by the entire branch (three for the Western Gondwana route and four
272 for the Northern Euramerica + East Asia route). If the dispersal path between two consecutive
273 ancestral states does not cross any of the five regions, then the number of formations in the
274 inhabited region is counted alone. Figure 1 illustrates an example of how this proxy is measured.
275 This results in the average number of formations present along the dispersal path (at geographic
276 region scale) for each branch in the phylogeny. As with stage-level formation counts, the regional- bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 11
277 level formation counts are weighted by the proportion that the branch length covers each geological
278 age. We hypothesize that dispersal rate will inversely correlate with regional-level formation count
279 because we expect that the lack of formations in intermediate regions will lead to inflated dispersal
280 rates. The ‘geo’ model will increase the dispersal rate along a branch to account for the geographic
281 variation observed when there is a lack of intermediate geographic fossil occurrences. This
282 hypothesis can be falsified if high dispersal rates are associated with larger average numbers of
283 formations along dispersal paths. Benton et al. (2013) provide a global sample of tetrapod-bearing
284 rock formations known for each geological age from the Middle Devonian through the Triassic.
285 We supplemented these lists with stratigraphic units known to produce sarcopterygian fossils
286 entered in the PBDB (collected on December 10th, 2018).
End Northern Southern Western Eastern East Period Epoch Age Time (Ma) Euramerica Euramerica Gondwana Gondwana Asia Total Permian Cisuralian Kungurian 272.95 10 44 11 0 0 65 Permian Cisuralian Artinskian 283.5 2 39 9 0 0 50 Permian Cisuralian Sakmarian 290.1 2 44 4 0 0 50 Permian Cisuralian Asselian 295 2 47 3 0 0 52 Pennsylvanian Late Gzhelian 298.9 1 42 0 0 0 43 Pennsylvanian Late Kasimovian 303.7 0 33 0 0 0 33 Pennsylvanian Middle Moscovian 307 0 16 0 0 0 16 Pennsylvanian Early Bashkirian 315.2 0 28 0 0 0 28 Mississippian Late Serpukhovian 323.2 0 16 0 0 0 16 Mississippian Middle Viséan 330.9 0 14 0 1 0 15 Mississippian Early Tournaisian 346.7 0 7 0 1 0 8 Devonian Late Famennian 358.9 1 9 1 5 1 17 Devonian Late Frasnian 372.2 1 11 0 3 2 17 Devonian Middle Givetian 382.7 1 8 1 4 1 15 Devonian Middle Eifelian 387.7 1 8 0 5 2 16 Devonian Early Emsian 393.3 2 5 0 7 3 17 Devonian Early Pragian 407.6 1 6 0 4 2 13 Devonian Early Lochkovian 410.8 1 3 0 3 1 8 Silurian Přídolí Přídolí 419.2 0 0 0 0 1 1 Silurian Ludlow Ludfordian 423 0 0 0 0 2 2 Silurian Ludlow Gorstian 425.6 0 0 0 0 2 2 287 Table 1: Regional- and stage-level (total) formation counts through time. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 12
288
289 To test for the effect of regional-level formation count bias on dispersal rate, we conducted
290 a non-parametric two-sample, upper-tailed Mann-Whitney U-test using the base package ‘stats’ in
291 R (R Core Team, 2018). This approach ranks all branches of the phylogeny by their regional-level
292 formation count and tests if the branches with lower dispersal rates rank higher on average than
293 branches with higher rates. We define “high” vs “low” dispersal rates based on whether or not they
294 are two standard deviations greater than the average rate across the tree. Due to the vast difference
295 in sample size between the two groups (“high rates”: n = 9, “low rates”: n = 111), we bootstrapped
296 the regional-level formation counts from each group with 100,000 replicates. From this bootstrap
297 analysis, we obtained a 95% confidence interval for the summed ranks of the branches with low
298 dispersal rates (n = 100,000 U-statistic values). The expected U-statistic is 499.5 given the null
299 hypothesis that only 50% of the regional-level formation counts along branches with low rates
9×111 300 rank higher than the formation counts with high rates (half of all possible combinations = ). 2
301 A 95% confidence interval of bootstrapped U-statistics that does not include the null expected U-
302 statistic is considered good evidence for higher mean dispersal rates along branches with lower
303 regional-level formation counts. The full dataset and code for the phylogeographic analyses can
304 be requested by email to the corresponding author.
305 Estimated ancestral states do not identify specific dispersal routes, so we conducted
306 sensitivity analyses to test if the dispersal route chosen for counting formations influenced our
307 results. We conceived of three scenarios for dispersal routes between Eastern Gondwana and
308 Southern Euramerica or vice versa: 1) a dispersal route through Western Gondwana; 2) a route
309 through Northern Euramerica and East Asia; and 3) a direct route between Eastern Gondwana and
310 Southern Euramerica. For the first scenario, we averaged the number of formations found in bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 13
311 Eastern and Western Gondwana and Southern Euramerica for a given time period. The second
312 scenario is similar to the first but included formation counts from Northern Euramerica and East
313 Asia in place of Western Gondwana. The third scenario only averaged formation counts from
314 Eastern Gondwana and Southern Euramerica.
315 3. Results
316 3.1. Supertree
317 Topological differences resulted among our supertree, the published source trees, and
318 Marjanović and Laurin's (2019) tree (Figure 2). In our tree, a polyphyletic “Megalichthyiformes”
319 is the basal-most tetrapodomorph group instead of Rhizodontida (Swartz, 2012; Zhu et al., 2017).
320 Canowindrids and rhizodontids formed an unexpected sister clade to Eotetrapodiformes. Clack et
321 al.’s (2017) five Tournaisian tetrapod taxa cluster together. Colosteidae is rootward of
322 Crassigyrinus. Caerorhachis is next to Baphetidae. Baphetidae moved crownward compared to
323 previous topologies (likely because of a small character sample size [Marjanović and Laurin,
324 2019]). Two crownward nodes are unresolved (polytomous). We retained Tungsenia and
325 Kenichthys as the oldest and second oldest tetrapodomorphs. Tristichopteridae, Elpistostegalia,
326 Stegocephalia, Aïstopoda, Whatcheeriidae, Colosteidae, Anthracosauria, Dendrerpetidae, and
327 Baphetidae remain monophyletic. Aïstopoda (Lethiscus and Coloraderpeton) fell rootward to
328 Tetrapoda as reported in Pardo et al. (2017; 2018). The average nRF distances quantify differences
329 in topology (see Supplementary Table 4). On average, there are 39.7% different or missing
330 bipartitions in the source trees compared to the supertree.
331 3.2. Phylogeography
332 We found overwhelming support for a variable rates model of geographic dispersal in early
333 tetrapodomorphs (BF = 632.3; Figure 3). The estimated rates across the three replicate runs are bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 14
334 consistent (out of 122 branches, only three had a median rate scalar with an absolute value
335 difference among the three runs greater than 3). All rate shifts that were two standard deviations
336 greater than the average dispersal rate were reconstructed dispersal events moving from East Asia
337 to Southern Euramerica, from Eastern Gondwana to Southern Euramerica, or Southern Euramerica
338 to Eastern Gondwana. The fastest estimated dispersal rate occurs along the branch leading to
339 Eotetrapodiformes, moving from Eastern Gondwana to Southern Euramerica (14.34x the average
340 rate). As Long et al. (2018) suggest, we find evidence for an East Asian origin for Tetrapodomorpha
341 but with moderate uncertainty (average estimate ± standard deviation of posterior distribution;
342 longitudeavg = 81.5° ± 10.1°, latitudeavg = −6.4° ± 8.5°). We also reconstruct an origin for
343 “Megalichthyiformes” that borderlines East Asia and Eastern Gondwana (longitudeavg = 107.2° ±
344 14.1°, latitudeavg = −22.6° ± 8.7°), along with an Eastern Gondwana origin for the clade uniting
345 “Canowindridae” and Rhizodontida (longitudeavg = 137.1° ± 8.2°, latitudeavg = −32.0° ± 4.7°). We
346 recover a Southern Euramerican origin for Eotetrapodiformes, consistent with previous studies
347 (longitudeavg = −12.5° ± 7.0°, latitudeavg = −19.4° ± 6.4°). A Southern Euramerican origin was also
348 found for Tristichopteridae (longitudeavg = −12.7° ± 6.9°, latitudeavg = −19.7° ± 6.3°) and
349 Elpistostegalia (longitudeavg = −12.3° ± 5.5°, latitudeavg = −13.5° ± 5.3°). As expected in a
350 phylogenetic comparative analysis, uncertainty in estimated node states increases toward the root.
351 However, despite the level of uncertainty within a single run, only three nodes have mean ancestral
352 state values that are greater than an absolute value of 5° among the replicate three runs.
353 We find good evidence that geographic sampling bias influences dispersal rate estimates,
354 regardless of the route used (95% CI: Western Gondwana route U = [800, 928]; Northern
355 Euramerica + East Asia route U = [832, 946]; direct route U = [729, 889]; no scenario includes the
356 null U = 499.5; Figure 4 and Supplementary Figures 12-13). A U-statistic considerably higher than bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 15
357 499.5 suggests that branches with high dispersal rates have lower regional-level formation counts,
358 on average, than branches with low rates. One can also interpret the null U-statistic of 499.5 as a
359 50% probability that a random branch with a low dispersal rate will rank higher in its regional-
360 level formation count than a random branch with a high dispersal rate. With bootstrapping, we are
361 95% confident that the probability of a random branch with a low dispersal rate having a higher
362 regional-level formation count than a random branch with a high rate is 72.97–88.99% for the
363 more conservative ‘direct route’ scenario. Under the more liberal ‘Northern Euramerica + East
364 Asia route’ scenario, the probabilities are 83.28–94.69%. In sum, branches with high dispersal
365 rates (two standard deviations greater than average) have a smaller number of recorded formations,
366 on average, along their reconstructed dispersal path.
367 Our results cannot be explained by a fossil record that is more complete through time (Pull of
368 the Recent). A regression model relating regional-level formation count to the minimum age of
369 each branch shows only a weak relationship (slope = -0.044, r2 = 0.1, P < 0.001). However, total
370 global (stage-level) formation count (which does not account for geographic variation) does show
371 potential bias from Pull of the Recent (slope = −0.3, r2 = 0.71, P < 0.0001). If dispersal rates are
372 biased by the increase in number of formations globally, we would also expect to see elevated
373 dispersal rates decrease toward the tips, but a regression model relating stretched branch lengths
374 with time is not supported (slope = −0.025, r2 = 0.006, P = 0.41).
375 4. Discussion
376 We expected to infer high dispersal rates for closely related taxa that are distributed across
377 the globe. Our results, unadjusted for geographic bias in the fossil record, confirm this notion.
378 However, we also find a compelling statistical association between high dispersal rates and a low
379 number of formations along dispersal paths—a patchy fossil record is driving inferences of high bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 16
380 dispersal rates. Although we did not test for a correlation between dispersal rate and previously
381 used proxies, such as valid taxon count and stage-level formation count, these proxies do not offer
382 clear predictions for explaining dispersal rate variation. High dispersal rate variation is inferred
383 when closely related taxa are geographically separate. For example, valid taxon count cannot
384 explain geographic rate variation because spatial information is lacking in this bias proxy and
385 because sister taxa are likely to have similar counts (these data are phylogenetically structured).
386 Stage-level formation counts will also not explain dispersal rate variation, particularly if high rate
387 variation exists within the same geological age. Assuming geological formations are evenly
388 exposed and sampled worldwide, low stage-level formation counts should yield geographically
389 variable fossil species and, therefore, drive high dispersal rate variation. However, formations are
390 not evenly exposed or recorded in geological/paleontological databases, including the PBDB. Our
391 formation count table demonstrates this bias (Table 1). Without geographic context, stage-level
392 formation count cannot distinguish between global and local regions. For example, the geological
393 ages that have the highest recorded number of formations are restricted to Southern Euramerica
394 where the majority of eotetrapodiform taxa have been discovered. The association between high
395 formation counts in specific regions and high paleobiodiversity in those regions is likely not a
396 coincidence and has a clear impact on how we interpret dispersal history. The earliest
397 tetrapodomorphs are known from China and Australia at geological ages where relatively few
398 formations are recorded outside of East Asia and Eastern Gondwana. The basal-most ancestral
399 state estimates reconstruct paleolocations in East Asia (not surprisingly). This inference
400 (hypothesis) is predicated on the lack of geological formations recorded outside of East Asia during
401 this time period. In addition, the majority of more crownward taxa and their reconstructed ancestral
402 states are located in North America and Europe at geological ages in which relatively fewer bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 17
403 formations are known elsewhere. This bias may heavily influence any conclusions made on the
404 location and habitat of the tetrapod water-land transition. Recently discovered taxa could help
405 mitigate this problem by increasing the power of taxon sampling (Heath et al., 2014), such as
406 Tutusius and Umzantsia from South Africa (Gess and Ahlberg, 2018). However, the current lack
407 of cladistic coding for these taxa excludes them from phylogeny-based analyses. The taxonomic
408 resolution of globally-occurring species, like Eusthenodon and Spodichthys, also impacts current
409 models of species dispersal history because of their relatively uniform distribution (Long et al.,
410 2018). Eusthenodon and Spodichthys represent possible cases where taxonomic resolution is too
411 coarse for phylogeographic analyses. Including these species inhibited our MCMC algorithms
412 from reaching convergence. Widely distributed cosmopolitan species that lack intermediate
413 geographic occurrences increase the uncertainty of parameter estimates within phylogeographic
414 models, as is the case here for these two species.
415 Phylogenetic studies on macroevolution also often fail to incorporate data from the fossil
416 record itself, such as trace fossil occurrences. Non-anatomical data often contribute to our
417 understanding of taxonomic originations, including chiridian (or digit-possessing)
418 tetrapodomorphs for which trace fossil evidence exists about 10 million years before the first
419 elpistostegalian body fossils (Niedźwiedzki et al., 2010). The inclusion of additional data from
420 trace fossils could radically alter our current models of species dispersal history. Finally, it is
421 important to note that the sampling bias proxies are also constrained by database curation biases.
422 Phylogenetic studies on macroevolutionary trends now regularly leverage public databases, such
423 as the PBDB, which allows larger and broader studies. It is unclear how patchy entries, on
424 taxonomic occurrences and geological formations, for example, interact with other biases inherent
425 in the fossil record. Caution is therefore warranted when these databases are mined, as is the case bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 18
426 here.
427 5. Conclusions
428 Phylogenetic studies on macroevolution have not previously incorporated geographic
429 context, which could influence a wide variety of analyses. We demonstrate here that
430 phylogeographic methods are influenced by geographic sampling variability. We develop a simple
431 sampling bias proxy that incorporates geographic information and show that it explains variation
432 in estimated dispersal rates. The majority of elevated dispersal rates are associated with large-scale
433 movements between major landmasses that have very few, if any, relevant geological formations
434 in between. Our analysis is also unlikely to be influenced by “Pull of the Recent”-like effects.
435 Although not the first supertree for early tetrapodomorphs (Ruta et al., 2003), this study presents the
436 first (to our knowledge) with branch lengths, making it useable for phylogenetic comparative
437 analyses. The new supertree comprises many of the major clades previously inferred, but also
438 recovers new ones that will be subject to scrutiny in future studies (discussed further in the
439 Supplementary Material). This supertree should be useful to researchers who aim to use
440 phylogenetic comparative methods to test hypotheses on the evolution of early tetrapodomorphs.
441 In sum, our study estimates ancestral geographical reconstructions consistent with previously
442 hypothesized dispersal patterns in early tetrapodomorphs. We also find that rates of dispersal are
443 strongly influenced by geographic sampling bias. We suggest that researchers incorporate this
444 proxy in phylogeny-based macroevolutionary studies that could be influenced by spatial
445 distribution of the fossil record.
446
447 Acknowledgements
448 We thank the MSU Macroevolution Lab, Jack Wilson, and Matt Lavin for helpful discussions, as bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 19
449 well as David Marjanović and John Long for their helpful reviews. We also thank Nathalie Bardet,
450 Eric Buffetaut, Annelise Folie, Emmanuel Gheerbrant, Alexandra Houssaye, and Michel Laurin
451 for the invitation to contribute to this special issue celebrating the life and accomplishments of
452 Jean-Claude Rage.
453
454 Funding
455 This research did not receive any specific grant from funding agencies in the public, commercial,
456 or not-for-profit sectors.
457
458 References
459 Ahlberg, P.E., 1998. Postcranial stem tetrapod remains from the Devonian of Scat Craig,
460 Morayshire, Scotland. Zool. J. Linn. Soc. 122, 99–141.
461 https://doi.org/10.1006/zjls.1997.0115
462 Alroy, J., Marshall, C.R., Bambach, R.K., Bezusko, K., Foote, M., Fürsich, F.T., Hansen, T.A.,
463 Holland, S.M., Ivany, L.C., Jablonski, D., Jacobs, D.K., Jones, D.C., Kosnik, M.A.,
464 Lidgard, S., Low, S., Miller, A.I., Novack-Gottshall, P.M., Olszewski, T.D., Patzkowsky,
465 M.E., Raup, D.M., Roy, K., Sepkoski, J.J., Sommers, M.G., Wagner, P.J., Webber, A., 2001.
466 Effects of sampling standardization on estimates of Phanerozoic marine diversification.
467 PNAS 98, 6261–6266. https://doi.org/10.1073/pnas.111144698
468 Benson, R.B.J., Butler, R.J., 2011. Uncovering the diversification history of marine tetrapods:
469 ecology influences the effect of geological sampling biases. Geological Society, London,
470 Special Publications 358, 191–208. https://doi.org/10.1144/SP358.13
471 Benson, R.B.J., Butler, R.J., Lindgren, J., Smith, A.S., 2010. Mesozoic marine tetrapod diversity: bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 20
472 mass extinctions and temporal heterogeneity in geological megabiases affecting vertebrates.
473 Proceedings of the Royal Society B: Biological Sciences 277, 829–834.
474 https://doi.org/10.1098/rspb.2009.1845
475 Benson, R.B.J., Upchurch, P., 2013. Diversity trends in the establishment of terrestrial vertebrate
476 ecosystems: Interactions between spatial and temporal sampling biases. Geology 41, 43–
477 46. https://doi.org/10.1130/G33543.1
478 Benton, M.J., Donoghue, P.C.J., Asher, R.J., Friedman, M., Near, T.J., Vinther, J., 2015.
479 Constraints on the timescale of animal evolutionary history. Palaeontol. Electron. 18, 1–
480 106. https://doi.org/10.26879/424
481 Benton, M.J., Ruta, M., Dunhill, A.M., Sakamoto, M., 2013. The first half of tetrapod evolution,
482 sampling proxies, and fossil record quality. Palaeogeography, Palaeoclimatology,
483 Palaeoecology 372, 18–41. https://doi.org/10.1016/j.palaeo.2012.09.005
484 Budd, G.E., Mann, R.P., 2018. History is written by the victors: The effect of the push of the past
485 on the fossil record. Evolution 72, 2276–2291. https://doi.org/10.1111/evo.13593
486 Clack, J.A., Bennett, C.E., Carpenter, D.K., Davies, S.J., Fraser, N.C., Kearsey, T.I., Marshall,
487 J.E.A., Millward, D., Otoo, B.K.A., Reeves, E.J., Ross, A.J., Ruta, M., Smithson, K.Z.,
488 Smithson, T.R., Walsh, S.A., 2017. Phylogenetic and environmental context of a
489 Tournaisian tetrapod fauna. Nat. Ecol. Evol. 1, 0002. https://doi.org/10.1038/s41559-016-
490 0002
491 Coates, M.I., Friedman, M., 2010. Litoptychus bryanti and characteristics of stem tetrapod
492 neurocrania, in: Elliot, D.K., Maisey, J.G., Yu, X., Miao, D. (Eds.), Morphology, Phylogeny
493 and Paleobiogeography of Fossil Fishes. Verlag Dr. Friedrich Pfeil, München, Germany,
494 pp. 389–416. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 21
495 Cope, E.D., 1868. Synopsis of the Extinct Batrachia of North America. Proceedings of the
496 Academy of Natural Sciences of Philadelphia. 20, 208–221.
497 Criscuolo, A., Berry, V., Douzery, E.J.P., Gascuel, O., 2006. SDM: A fast distance-based approach
498 for (super)tree building in phylogenomics. Syst. Biol. 55, 740–755.
499 https://doi.org/10.1080/10635150600969872
500 Criscuolo, A., Gascuel, O., 2008. Fast NJ-like algorithms to deal with incomplete distance matrices.
501 BMC Bioinformatics 9, 166. https://doi.org/10.1186/1471-2105-9-166
502 Didier, G., Fau, M., Laurin, M., 2017. Likelihood of tree topologies with fossils and diversification
503 rate estimation. Syst. Biol. 66, 964–987. https://doi.org/10.1093/sysbio/syx045
504 Didier, G., Laurin, M., 2018. Exact distribution of divergence times from fossil ages and topologies.
505 bioRxiv 490003. https://doi.org/10.1101/490003
506 Didier, G., Royer-Carenzi, M., Laurin, M., 2012. The reconstructed evolutionary process with the
507 fossil record. J. Theor. Biol. 315, 26–37. https://doi.org/10.1016/j.jtbi.2012.08.046
508 Dunhill, A.M., Benton, M.J., Newell, A.J., Twitchett, R.J., 2013. Completeness of the fossil record
509 and the validity of sampling proxies: a case study from the Triassic of England and Wales.
510 Journal of the Geological Society 170, 291–300. https://doi.org/10.1144/jgs2012-025
511 Dunhill, A.M., Benton, M.J., Twitchett, R.J., Newell, A.J., 2014a. Testing the fossil record:
512 Sampling proxies and scaling in the British Triassic–Jurassic. Palaeogeography,
513 Palaeoclimatology, Palaeoecology 404, 1–11. https://doi.org/10.1016/j.palaeo.2014.03.026
514 Dunhill, A.M., Hannisdal, B., Benton, M.J., 2014b. Disentangling rock record bias and common-
515 cause from redundancy in the British fossil record. Nature Communications 5, 4818.
516 https://doi.org/10.1038/ncomms5818
517 Foote, M., 2003. Origination and Extinction through the Phanerozoic: A New Approach. The bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 22
518 Journal of Geology 111, 125–148. https://doi.org/10.1086/345841
519 Friedman, M., Coates, M.I., Anderson, P., 2007. First discovery of a primitive coelacanth fin fills
520 a major gap in the evolution of lobed fins and limbs. Evol. Dev. 9, 329–337.
521 https://doi.org/10.1111/j.1525-142X.2007.00169.x
522 Gascuel, O., 1997. Concerning the NJ algorithm and its unweighted version, UNJ, in: Roberts, F.,
523 Rzhetsky, A. (Eds.), Mathematical Hierarchies and Biology. American Mathematical Soc.,
524 Providence, RI, pp. 149–170.
525 Gauthier, J., Cannatella, D., de Queiroz, K., Kluge, A.G., Rowe, T., 1989. Tetrapod phylogeny, in:
526 Fernholm, B., Bremer, K., Jörnvall, H. (Eds.), The Hierarchy of Life. Elsevier Science
527 Publishers B. V. (Biomedical Division), Amsterdam, Netherlands.
528 Gavryushkina, A., Welch, D., Stadler, T., Drummond, A.J., 2014. Bayesian Inference of Sampled
529 Ancestor Trees for Epidemiology and Fossil Calibration. PLOS Comput Biol 10, e1003919.
530 https://doi.org/10.1371/journal.pcbi.1003919
531 Gess, R., Ahlberg, P.E., 2018. A tetrapod fauna from within the Devonian Antarctic Circle. Science
532 360, 1120–1124. https://doi.org/10.1126/science.aaq1645
533 Heath, T.A., Huelsenbeck, J.P., Stadler, T., 2014. The fossilized birth–death process for coherent
534 calibration of divergence-time estimates. PNAS 111, E2957–E2966.
535 https://doi.org/10.1073/pnas.1319091111
536 Jablonski, D., Roy, K., Valentine, J.W., Price, R.M., Anderson, P.S., 2003. The Impact of the Pull
537 of the Recent on the History of Marine Diversity. Science 300, 1133–1135.
538 https://doi.org/10.1126/science.1083246
539 Koch, C.F., 1978. Bias in the Published Fossil Record. Paleobiology 4, 367–372.
540 Lakner, C., van der Mark, P., Huelsenbeck, J.P., Larget, B., Ronquist, F., 2008. Efficiency of bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 23
541 Markov chain Monte Carlo tree proposals in Bayesian phylogenetics. Syst. Biol. 57, 86–
542 103. https://doi.org/10.1080/10635150801886156
543 Laurin, M., 1998. The importance of global parsimony and historical bias in understanding
544 tetrapod evolution. Part I. Systematics, middle ear evolution and jaw suspension. Ann. Sci.
545 Nat. Zoo. 19, 1–42. https://doi.org/10.1016/S0003-4339(98)80132-9
546 Lloyd, G.T., 2012. A refined modelling approach to assess the influence of sampling on
547 palaeobiodiversity curves: new support for declining Cretaceous dinosaur richness. Biol.
548 Lett. 8, 123–126. https://doi.org/10.1098/rsbl.2011.0210
549 Long, J.A., Clement, A.M., Choo, B., 2018. New insights into the origins and radiation of the mid-
550 Palaeozoic Gondwanan stem tetrapods. Earth and Environmental Science Transactions of
551 The Royal Society of Edinburgh 1–17. https://doi.org/10.1017/S1755691018000750
552 Marjanović, D., Laurin, M., 2019. Phylogeny of Paleozoic limbed vertebrates reassessed through
553 revision and expansion of the largest published relevant data matrix. PeerJ 6, e5565.
554 https://doi.org/10.7717/peerj.5565
555 Marshall, J.E.A., Reeves, E.J., Bennett, C.E., Davies, S.J., Kearsey, T.I., Millward, D., Smithson,
556 T.R., Browne, M.A.E., 2019. Reinterpreting the age of the uppermost “Old Red Sandstone”
557 and Early Carboniferous in Scotland. Earth Env. Sci. T. R. So. 109, 265–278.
558 https://doi.org/10.1017/S1755691018000968
559 Niedźwiedzki, G., Szrek, P., Narkiewicz, K., Narkiewicz, M., Ahlberg, P.E., 2010. Tetrapod
560 trackways from the early Middle Devonian period of Poland. Nature 463, 43–48.
561 https://doi.org/10.1038/nature08623
562 O’Donovan, C., Meade, A., Venditti, C., 2018. Dinosaurs reveal the geographical signature of an
563 evolutionary radiation. Nature Ecology & Evolution 2, 452. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 24
564 https://doi.org/10.1038/s41559-017-0454-6
565 Pardo, J.D., Szostakiwskyj, M., Ahlberg, P.E., Anderson, J.S., 2017. Hidden morphological
566 diversity among early tetrapods. Nature 546, 642–645.
567 https://doi.org/10.1038/nature22966
568 R Core Team, 2018. R: A Language and Environment for Statistical Computing. R Foundation for
569 Statistical Computing, Vienna, Austria.
570 Rambaut, A., Drummond, A.J., Xie, D., Baele, G., Suchard, M.A., 2018. Posterior Summarization
571 in Bayesian Phylogenetics Using Tracer 1.7. Syst Biol 67, 901–904.
572 https://doi.org/10.1093/sysbio/syy032
573 Raup, D.M., Boyajian, G.E., 1988. Patterns of Generic Extinction in the Fossil Record.
574 Paleobiology 14, 109–125.
575 Revell, L.J., 2012. phytools: an R package for phylogenetic comparative biology (and other things).
576 Methods in Ecology and Evolution 3, 217–223. https://doi.org/10.1111/j.2041-
577 210X.2011.00169.x
578 Robinson, D.F., Foulds, L.R., 1981. Comparison of phylogenetic trees. Math. Biosci. 53, 131–147.
579 https://doi.org/10.1016/0025-5564(81)90043-2
580 Ronquist, Fredrik, Klopfstein, S., Vilhelmsen, L., Schulmeister, S., Murray, D.L., Rasnitsyn, A.P.,
581 2012. A total-evidence approach to dating with fossils, applied to the early radiation of the
582 Hymenoptera. Syst. Biol. 61, 973–999. https://doi.org/10.1093/sysbio/sys058
583 Ronquist, F., Teslenko, M., van der Mark, P., Ayres, D., Darling, A., Höhna, S., Larget, B., Liu, L.,
584 Suchard, M.A., Huelsenbeck, J.P., 2012. MrBayes 3.2: Efficient Bayesian phylogenetic
585 inference and model choice across a large model space. Systematic Biology.
586 Rothkugel, S., Varela, S., 2015. paleoMap: An R-package for getting and using paleontological bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 25
587 maps.
588 Sakamoto, M., Benton, M.J., Venditti, C., 2016a. Dinosaurs in decline tens of millions of years
589 before their final extinction. PNAS 113, 5036–5040.
590 https://doi.org/10.1073/pnas.1521478113
591 Sakamoto, M., Venditti, C., Benton, M.J., 2016b. ‘Residual diversity estimates’ do not correct for
592 sampling bias in palaeodiversity data. Methods Ecol Evol n/a-n/a.
593 https://doi.org/10.1111/2041-210X.12666
594 Schliep, K.P., 2011. phangorn: Phylogenetic analysis in R. Bioinformatics 27, 592–593.
595 https://doi.org/10.1093/bioinformatics/btq706
596 Signor, P.W., Lipps, J.H., 1982. Sampling bias, gradual extinction patterns, and catastrophes in the
597 fossil record. Geological Society of America Special Publication 190, 291–296.
598 Stadler, T., 2010. Sampling-through-time in birth-death trees. J. Theor. Biol. 267, 396–404.
599 https://doi.org/10.1016/j.jtbi.2010.09.010
600 Swartz, B., 2012. A marine stem-tetrapod from the Devonian of Western North America. PLoS
601 ONE 7, e33683. https://doi.org/10.1371/journal.pone.0033683
602 Tennant, J.P., Mannion, P.D., Upchurch, P., 2016a. Environmental drivers of crocodyliform
603 extinction across the Jurassic/Cretaceous transition. Proc. R. Soc. B 283, 20152840.
604 https://doi.org/10.1098/rspb.2015.2840
605 Tennant, J.P., Mannion, P.D., Upchurch, P., 2016b. Sea level regulated tetrapod diversity dynamics
606 through the Jurassic/Cretaceous interval. Nature Communications 7, 12737.
607 https://doi.org/10.1038/ncomms12737
608 Venditti, C., Meade, A., Pagel, M., 2011. Multiple routes to mammalian diversity. Nature 479,
609 393–396. https://doi.org/10.1038/nature10516 bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 26
610 Xie, W., Lewis, P.O., Fan, Y., Kuo, L., Chen, M.-H., 2011. Improving Marginal Likelihood
611 Estimation for Bayesian Phylogenetic Model Selection. Syst Biol 60, 150–160.
612 https://doi.org/10.1093/sysbio/syq085
613 Zhang, C., Stadler, T., Klopfstein, S., Heath, T.A., Ronquist, F., 2016. Total-Evidence Dating under
614 the Fossilized Birth–Death Process. Syst Biol 65, 228–249.
615 https://doi.org/10.1093/sysbio/syv080
616 Zhu, M., Ahlberg, P.E., Zhao, W.-J., Jia, L.-T., 2017. A Devonian tetrapod-like fish reveals
617 substantial parallelism in stem tetrapod evolution. Nat. Ecol. Evol. 1, 1470–1476.
618 https://doi.org/10.1038/s41559-017-0293-5
619
620 Figure Captions
621 1. Example of how the regional-level formation count proxy is calculated. A) Five major
622 geographic regions are highlighted by color in the Devonian map. Red arrows represent a branch-
623 specific dispersal path to species A, beginning in Southern Euramerica and ending in Eastern
624 Gondwana. The blue arrow represents the dispersal path to species B. B) The phylogeny of species
625 A and B scaled by time, with equal branch lengths to both species, and colored to represent the
626 rate of dispersal (red is fast, blue is slow). For every branch of the tree, the number of formations
627 is counted for every region and for each geological age covered by the dispersal pathway. It is then
628 weighted by the number of geological ages and geographic regions covered. Under the Western
629 Gondwana route scenario, the branch to species A covers three geographic regions, while the
630 branch to species B only covers one. Assuming both branches cover only one geological age, the
631 high dispersal rate for species A can be explained by the lack of recorded geological formations in
632 Western Gondwana. C) A line plot of the formation counts through time, colored by geographic bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 27
633 region according to the Devonian map above, shows temporal and geographic variability.
634
635 2. The time-scaled tetrapodomorph supertree. Taxonomic groups in quotes are not monophyletic.
636 Here, Glyptolepis, a dipnomorph, is the outgroup. We downloaded the silhouettes from
637 phylopic.org: Eucritta and Greererpeton by Dmitry Bogdanov (vectorized by Michael Keesey),
638 Eusthenopteron by Steve Coombs (vectorized by Michael Keesey), and Gogonasus and Tiktaalik
639 by Nobu Tamura (CC BY-SA 3.0).
640
641 3. A) Trimmed tetrapodomorph phylogeny with mapped rates of dispersal. Cooler (bluish) colors
642 represent slower rates and warmer (reddish) colors represent faster rates. B) Non-eotetrapodiform
643 (left in blue) and eotetrapodiform (right in green) trees and taxon paleolocations plotted on a map
644 of the Middle Devonian. Transparent polygons illustrate broad geographic regions of sampled taxa
645 in Southern Euramerica, Eastern Gondwana, and East Asia. Numbers show the total number of
646 geological formations recorded from each major geographic region (Eastern Gondwana and East
647 Asia combined). Colored circles show average paleolocations of major clades estimated by the
648 ‘geo’ model and indicated in the tree above. Red circle: Tetrapodomorpha, orange:
649 “Megalichthyiformes”, yellow: “Canowindridae” + Rhizodontidae, green: Tristichopteridae, and
650 blue: Elpistostegalia. Phylogeny with mapped dispersal rates was produced in BayesTrees
651 (http://www.evolution.rdg.ac.uk/BayesTrees.html). Middle Devonian tree and paleolocation plots
652 were made using the ‘phylo-to-map’ function in the R package, phytools (Revell, 2012). Middle
653 Devonian map was sourced from the R package, paleoMap (Rothkugel and Varela, 2015).
654 Tetrapodomorph silhouettes were sourced from phylopic.org: Eucritta by Dmitry Bogdanov
655 (vectorized by T. Michael Keesey), Osteolepis by Nobu Tamura, and Acanthostega by Mateus Zica. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. 28
656
657 4. A) Scatter-plot of the average dispersal rates over the regional-level formation counts for each
658 branch of the phylogeny, using the Northern Euramerica + East Asia route scenario. Points colored
659 by the dispersal rate being above (red) or below (blue) two standard deviations greater than the
660 average rate across the tree. B) Histogram of the bootstrapped U-statistics. Values outside of the
661 95% confidence interval are grayed out. The median and null expected U-statistics are indicated
662 by the red and blue dotted lines, respectively. The null expected U-statistic is based on the null
663 hypothesis that 50% of the branches with low dispersal rates will have a greater regional-level
664 formation count than branches with higher rates. Rejecting the null hypothesis suggests that
665 estimated dispersal rates are biased and correlate with regional-level formation count. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license. bioRxiv preprint doi: https://doi.org/10.1101/726786; this version posted August 6, 2019. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-NC-ND 4.0 International license.